In a recent essay, AI researcher Nathan Lambert argues that the field is heading for rapid progress, but not toward general superintelligence. The acceleration, he says, will come from infrastructure and engineering improvements — models will outperform humans at distributed GPU engineering within a few years — rather than from a fundamental change in the nature of the models themselves.
Much of the near-term gain will come from scaling inference-time compute and squeezing efficiency out of existing tools. Lambert expects AI agents to optimize the training and inference stack end-to-end, driving the effective cost of model intelligence down at a near-exponential rate. He predicts that pretraining research on architecture and data selection could be automated within two to three years.
Lambert sees this as a shift in the field's bottleneck, away from engineering and back toward ideas, similar to the pre-deep-learning era. He also expects demand for agentic models to keep rising — a Jevons paradox — and points to Meta's Muse agent as an early example. In science, he argues, models are already superhuman at crawling literature and connecting sparse research communities, which could accelerate cross-subfield discovery.
Not everything is rosy. Lambert notes that much of the RL data sold by a booming industry is low quality, though he says the problems are fixable and leading labs see clear returns on the data they buy. His overall message is grounded: rapid, compounding progress in engineering and efficiency, but no reason to buy into takeoff scenarios.